Introducing Decagon Labs

Today, over 80% of all model traffic at Decagon runs on models we trained ourselves. They outperform the best foundation models on our real-world use cases. 

The team behind that work is Decagon Labs.

Why we built our own models

When we started Decagon, we relied on the same foundation models as everyone else. They are very powerful, but they aren’t specifically built for what we do.

Enterprise customer experience (CX) demands precision, speed, and reliability at a level that general-purpose models aren't optimized for. As the major labs push toward broader reasoning capabilities, the gap between what they prioritize and what enterprise CX actually requires has only widened.

Our answer was to build a different architecture entirely: a network of specialized models, each responsible for a distinct function—identifying the end of user speech, executing workflows, detecting hallucinations, and more. Post-training models specifically for each role lets us hit the low latency customers expect and the high accuracy they depend on. We wrote about this in depth in our post on fine-tuning AI agents.

That architecture demanded a team capable of building it.

Meet Decagon Labs

Decagon Labs is the research and agent orchestration arm of our engineering team. We focus on frontier AI research with a singular goal: make the Decagon product better

That means developing state-of-the-art training techniques, new model architectures, and rigorous evaluation methods, all in direct service of the agents our customers deploy. We've built the infrastructure to support it: human annotators, end-to-end evaluation pipelines, training and inference platforms, and a research team that ships to production.

We're publishing four deep dives alongside this post that give a window into our current work:

• Why off-policy training isn't enough: SFT, RL, and the limits of imitation
• Optimizing GEPA for production: A test-driven approach to prompt engineering
• Bayesian VAD: Fixing turn detection in production voice pipelines
• Designing low-latency AI agents through reranker optimization

Research that ships

Our thesis is that the future of enterprise CX isn't a bigger general-purpose model, but a system of smarter, more specialized models.

The major foundation labs are solving massive, important problems: general reasoning, coding, scientific discovery. Our focus is narrower and more applied: building the best models for delivering concierge customer experiences. That specialization lets us move faster and build deeper in the areas that matter most.

Every member of the Decagon Labs team has a direct line from their work to production impact. The models they build power millions of customer interactions. The architectures they design get deployed to the largest brands in the world. Research here is operational rather than theoretical.

What's next

Our current focus is building frontier post-training techniques and agent architectures that push the boundaries of what our real-world agents can do. We're also prioritizing our voice agents, an area where we believe there's enormous room to improve on the off-the-shelf components available today.

As we scale, we're expanding the scope of our proprietary model stack and deepening our investments in evaluation, training infrastructure, and the research team itself. You'll be hearing a lot more from us through technical blog posts, conference talks, and more.

If you're an engineer or researcher who wants to do frontier work that ships to production and impacts real customers at scale, we're hiring. Come build with us.